Store a durable memory (fact/preference/pattern/decision/outcome) that persists across sessions. Every write is scanned for secrets (16+ patterns — OpenAI/Stripe/AWS/GitHub/Slack/Google keys, private keys, connection strings, JWTs): by default detected secrets are redacted to [REDACTED] before storage, or the write is rejected if auto-redaction is disabled. Category and entity are auto-extracted when omitted, a local embedding is generated, and the content is deduplicated against existing memories. Returns: the memory id plus an outcome — "created" (new), "merged" (0.92–0.95 cosine to an existing memory; content/tags/confidence folded into it), or "duplicate" (≥0.95 cosine; not stored unless force:true). Use when you learn something worth remembering about the user, project, setup, or workflow; recall with engram_recall, delete with engram_forget.

Retrieve memories relevant to a query, ranked by a hybrid score. Embeds the query, gathers candidates (FTS5 top-20 plus in-namespace embeddings, optionally time-filtered), and scores each by similarity×0.45 + recency×0.15 + confidence×0.15 + access×0.05 + feedback×0.10 + a 0.1 FTS boost, then filters by category/threshold and returns the top results. If embedding generation fails it falls back to FTS-only search. Reading a memory bumps its last_accessed and access_count. Returns: an array of memory objects — each with id, content, category, entity, confidence, namespace, tags, timestamps, score, and scoreBreakdown — or an empty array if nothing clears the threshold (with a time_filter, the array also carries timeRange metadata). Use at session start or to look up a specific fact; prefer engram_context when you want a ready-to-inject block instead of raw results.

Permanently delete one memory by id. Irreversible — also removes that memory's feedback rows. Returns: whether a memory with the given id was found and deleted; reports not-found without error if the id doesn't exist. Use when a memory is wrong, outdated, or the user asks you to forget it. If you're unsure, downvote with engram_feedback (helpful:false) instead of deleting.

Record a helpful/unhelpful vote on a recalled memory to tune future ranking. Updates the memory's aggregated feedback_score (−1 to 1), which feeds the recall score (weight 0.10); after 5+ votes it may auto-adjust the memory's confidence (strongly negative lowers it, strongly positive raises it). Returns: the updated feedback stats for that memory. Call right after acting on a memory from engram_recall to close the learning loop; to remove a bad memory outright, use engram_forget instead.

Build a single pre-formatted context block from relevant memories, ready to inject into a system prompt at session start. With a query it selects semantically relevant memories; with no query it returns the top memories by access frequency and recency. The block is rendered in the requested format and truncated to fit max_tokens. Returns: one formatted string (not an array) — contrast with engram_recall, which returns raw scored memory objects. Use when you want drop-in context text; use engram_recall when you need structured results to reason over.

Report Engram health and statistics. Read-only and parameter-free. Returns: memory counts by category and namespace, embedding-model status (name, cached/loaded state, size), the database location, and key config (default namespace, recall limit, confidence threshold, secret-detection on/off). Use as a diagnostics/health check — to confirm the model is loaded and see how many memories exist — before relying on recall.